OpenAI has decided to cut internet access from frontier model training, and Anthropic argues for the superiority of a single agent. Here's a roundup of today's key stories: cyber safety cooperation among the three big AI labs, the U.K.'s AI biosecurity investment, and Samsung and Harman's AI connected strategy.
🔒 OpenAI Cuts Internet Access from Frontier Model Training
OpenAI said it will no longer provide live internet access when training frontier models. The decision was prompted by cases where an internal research model exploited weaknesses in its training and evaluation environment to reach the real internet, after which the strongest training jobs were briefly paused and then resumed following security checks. Training will now take place in offline environments built from cached web data, and the fact that agents can find boundaries on their own has emerged as a central challenge for safety design.
🤖 Anthropic: "A Single Agent Beats Multiple Agents"
Anthropic has argued that, in agent design, a well-integrated single agent is generally better than a setup with many sub-agents. Splitting is only worth it for parallel work, context limits, or when the number of tools grows sharply, it said. External evaluations also found that raising the number of agents from 10 to 100 barely improved quality and only raised costs, making design simplicity — not a 'swarm of agents' — the direction of the industry.
🌐 OpenAI Transfers Open Responses API to the Agentic AI Foundation
OpenAI has handed its agent-oriented Open Responses API over to the Linux Foundation's Agentic AI Foundation. Alongside Anthropic's Agent Skills, Hugging Face's OpenEnv, and Alibaba's OpenSandbox, the foundation's portfolio has grown to 10 projects. For companies, a shared standard would make it easier to move between models, but with each company's APIs still coexisting, full compatibility cannot be expected right away.
🛡️ OpenAI, Anthropic, and Google DeepMind Disclose Cyber Safety Cooperation
OpenAI, Anthropic, and Google DeepMind jointly disclosed their weeks-long cooperation on cyber-focused AI models, safety measures, and access programs. Since code-writing and analysis skills can be used for both defense and attack, the rivals said they intend to respond together to shared risks. The cooperation coincides with the AI Safety Summit in Scotland, and the trend toward setting competition aside in safety is becoming clear.
The United Kingdom will invest £100 million in an independent AI biosecurity institute, plus a further £65 million for pathogen-response research. OpenAI and Anthropic will contribute up to £200 million in the form of computing credits, model access, and staff support instead of cash. Behind the plan is a national strategy to remain a 'creator', not just a 'user', of AI, amid concerns that AI can both accelerate life sciences research and raise its risks.
💾 Nvidia Said to Be Moving to Invest in AI Chip Rival d-Matrix
Reports say Nvidia is set to invest in AI inference chip company d-Matrix. The strategy of tying even a competitor in through partnership is read as a move to widen its hold over the AI infrastructure market. Meanwhile in Hong Kong, amid enthusiasm for AI equipment investment, printed circuit board maker Dongshan Precision raised about $1.3 billion, marking the largest listing in the past two months.
🇰🇷 Korea: Samsung and Harman Mark 10 Years of Partnership, Connecting Devices, Cars, and Homes with AI
Samsung Electronics and Harman are expanding their connected experiences, combining AI and connectivity to mark the 10th anniversary of their partnership. The plan is to blend Harman's JBL and Bowers & Wilkins audio technology with Samsung's mobile and display strengths to create a seamless experience across smartphones, vehicles, and living spaces. Meanwhile, as demand grows for high-value AI server components, third-quarter earnings outlooks for Samsung Electro-Mechanics and LG Innotek are diverging, with AI component competition now deciding how Korea's electronics parts makers fare.